Machine learning identifies a DNA repair-related risk model for bladder cancer and functionally characterizes ZWINT as a potential oncogenic factor

Abstract

Background

Bladder cancer (BCa) is characterized by substantial molecular and clinical heterogeneity, hindering accurate prognostic assessment and treatment stratification. This study aimed to develop a machine learning-based prognostic model based on DNA repair-related genes and to investigate the functional relevance of representative genes in BCa.

Methods

To address this, we constructed a DNA repair-related risk score (DRRS) using 101 machine learning algorithms based on transcriptomic profiles from BCa cohorts. The prognostic value, immune relevance, and therapeutic implications of DRRS were systematically evaluated. ZW10 interacting kinetochore protein (ZWINT), selected as a representative gene within the DRRS, was subjected to functional characterization through in vitro experiments.

Results

DRRS effectively stratified BCa patients into high- and low-risk groups with distinct survival outcomes, immune microenvironment features, and associations with responses to chemotherapy and immunotherapy. Functional assays revealed that ZWINT knockdown suppressed cell proliferation, migration, and invasion, induced apoptosis, and caused G1 phase cell cycle arrest. Transcriptomic analysis and Western blot validation demonstrated increased phosphorylation of p65 following ZWINT knockdown, indicating activation of the NF-κB signaling pathway, potentially as part of a feedback mechanism associated with apoptosis.

Conclusions

This study establishes a machine learning-based DNA repair signature for BCa and functionally validates ZWINT as a potential oncogenic driver. The findings provide new insight into BCa progression and highlight DRRS and ZWINT as promising biomarkers and therapeutic targets for treatment stratification.

Introduction

Bladder cancer (BCa) is a serious malignancy with a notably high incidence rate around the world. According to epidemiological data, it ranks as the tenth most common cancer globally and is the second most frequent urological malignancy (1,2). Clinically, BCa is categorized into two primary subtypes based on the extent of tumor invasion: non-muscle-invasive BCa (NMIBC) and muscle-invasive BCa (MIBC). While NMIBC is associated with a relatively low risk of metastasis and mortality, recurrence is common, with 15–20% of patients potentially progressing to the more invasive MIBC (3). In contrast, MIBC is characterized by a poorer prognosis and presents greater challenges for treatment. In recent years, substantial advancements have been achieved in both the prevention and treatment of BCa (4), particularly through the application of immunotherapy involving immune checkpoint blockers (ICBs) (5). While immunotherapy has revolutionized cancer treatment, its effectiveness differs significantly across individuals, and traditional staging approaches fall short in predicting responses. Models leveraging biomarkers and complex data may help refine patient stratification and improve therapeutic assessment.

DNA repair is an essential mechanism for preserving genomic stability, addressing various types of DNA damage through distinct pathways, such as base excision repair (BER), nucleotide excision repair (NER), mismatch repair (MMR), homologous recombination (HR), and non-homologous end joining (NHEJ) (6,7). These processes rely heavily on DNA repair-related genes, of which its functional loss or aberrations can trigger DNA damage accumulation and genomic instability, thereby driving tumor formation and advancement (8). For instance, alterations in the BRCA1/2 genes are strongly associated with increased susceptibility to breast and ovarian cancers (9,10). In addition, DNA repair gene defects raise tumor mutation burdens and enhance immunogenicity. For example, microsatellite instability-high (MSI-H) colon cancer, characterized by a high mutation load and abundant neoantigens, frequently demonstrates significant sensitivity to immune checkpoint inhibitors (ICIs) (11,12). While previous bioinformatics studies have explored the potential roles of DNA repair-related genes in BCa, the lack of experimental validation restricts their applicability (13). Building on these findings, we applied a machine learning-based framework to refine computational predictions and provide experimental support, thereby offering novel insights into the biological mechanisms and therapeutic potential of DNA repair-related genes in BCa.

In this study, we utilized 101 machine learning algorithms to construct a reliable DNA repair-related risk score (DRRS). We further analyzed the distinctions between the two risk groups regarding immune infiltration, immunotherapy responsiveness, and chemotherapy outcomes. Furthermore, the functional role of ZW10 interacting kinetochore protein (ZWINT) in BCa was experimentally validated. We present this article in accordance with the MDAR and TRIPOD reporting checklists (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-710/rc).

Methods

Dataset collection

The microarray datasets GSE32894, GSE3167, GSE120736, GSE13507, GSE5287, GSE169455, and GSE48075 and the sequencing dataset GSE188715 were retrieved from the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/). Bulk RNA sequencing (RNA-seq) data and clinicopathological features of bladder urothelial carcinoma (BLCA) were obtained from The Cancer Genome Atlas (TCGA; https://portal.gdc.cancer.gov/) in June 2024 (14). For analyses requiring survival information, samples lacking survival data were excluded; otherwise, all available samples were retained. A total of 150 DNA repair-related genes were downloaded from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb/) for analysis (15). The information of all datasets is shown in Table S1.

Identification of differentially expressed genes (DEGs)

DEGs were identified using the edgeR package. Probe sets without corresponding gene symbols were excluded. DEGs were screened based on the criteria of adjusted P value <0.05 and |log₂fold change (FC)| ≥0.3. To visualize these DEGs, heatmaps and volcano plots were created using the R software. The overlap of DEGs was identified through a Venn diagram, generated using an online tool (https://bioinfogp.cnb.csic.es/tools/venny).

Machine learning algorithms

The machine learning methodology used to construct the DRRS followed a previously published framework described by Liu et al. (16), is summarized in Figure S1. In brief, we combined 10 distinct machine learning algorithms and evaluated 101 different algorithmic combinations. These algorithms included survival support vector machine (survivalSVM), least absolute shrinkage and selection operator (Lasso), elastic net (Enet), Ridge, random survival forest (RSF), generalized boosted regression modeling (GBM), stepwise Cox, CoxBoost, partial least squares regression for Cox (plsRcox), and supervised principal components (SuperPC). Our approach began with the identification of prognostic variables through univariate Cox regression modeling. Prediction models were then developed using the TCGA-BLCA cohort, with the GSE32894 dataset serving as the validation cohort. To select the best-performing model, we calculated the Harrell’s concordance index (C-index). The model with the highest average C-index across both cohorts was defined as the optimal model. The DRRS for each patient was calculated as the weighted sum of gene expression levels, with weights assigned based on the corresponding regression coefficients:

$$Riskscore ={{\displaystyle \sum }}_{k=1}^{n}Coef\left(k\right) \ast Exp\left(k\right)$$ [1]

where n represents the number of genes included in the model, Coef(k) denotes the regression coefficient, and Exp(k) corresponds to the expression level of each gene. Subsequently, we classified BLCA patients into high- and low-DRRS groups using the median score of each cohort. Prognostic significance was assessed using Kaplan-Meier survival curves, and the prognostic utility of the DRRS was further evaluated using Receiver operating characteristic (ROC) curves. Finally, the risk score was validated as an independent prognostic factor through Cox regression analysis.

Assessment of the immune microenvironment

To comprehensively assess the extent of immune infiltration, we employed a variety of bioinformatic algorithms, including Tumor Immune Estimation Resource (TIMER), Cell-Type Identification by Estimating Relative Subpopulations of RNA Transcripts (CIBERSORT), CIBERSORT absolute mode (CIBERSORT-ABS), single-sample gene set enrichment analysis (ssGSEA), Microenvironment Cell Populations-counter (MCP-counter), QUANTISEQ, xCell, and Estimation of Stromal and Immune cells in Malignant Tumor tissues using Expression data (ESTIMATE). Furthermore, we analyzed the expression patterns of 79 immune checkpoint genes using the Wilcoxon signed-rank test. Additionally, we utilized the Tumor Immune Dysfunction and Exclusion (TIDE) (http://tide.dfci.harvard.edu/), which stands for Tumor Immune Dysfunction and Rejection, to evaluate responsiveness to immunotherapy. The analysis was further validated using the real-world IMvigor210 immunotherapy cohort. Furthermore, the GSE169455 dataset was used to validate the responsiveness to chemotherapy.

Functional enrichment analysis

Genes exhibiting significant differential expression between the control and ZWINT knockdown groups were identified based on the thresholds of adjusted P value <0.05 and |log₂FC| >1. The R packages ‘clusterProfiler’ and ‘org.Hs.eg.db’ were employed to analyze biological functions based on Gene Ontology (GO) and to perform pathway enrichment analysis for the Kyoto Encyclopedia of Genes and Genomes (KEGG). In addition, gene set enrichment analysis (GSEA) was conducted to identify enriched pathways among DEGs between the ZWINT knockdown and control groups.

Immunohistochemical (IHC) staining

A tissue microarray containing nine paired BCa tissues and adjacent normal tissues was used for immunohistochemistry analysis. All patients were newly diagnosed BCa cases without prior treatment. Following rehydration, the tissue specimens were treated with ethylenediaminetetraacetic acid (EDTA) buffer (pH 9.0) under high pressure for 10 minutes and subsequently immersed in 3% hydrogen peroxide at room temperature for 10 minutes. Tissues were incubated with anti-ZWINT overnight, then treated with a biotin-labeled goat anti-rabbit secondary antibody for 20 minutes at room temperature, followed by 5 minutes of 3,3'-diaminobenzidine (DAB) staining and 30 seconds of hematoxylin staining (ZSG Bio, Beijing, China).

Cell lines and cell culture

The human BCa cell lines T24 and UM-UC-3 were obtained from the Cell Bank of the Shanghai Institute of Cell Biology (Shanghai, China). Both cell lines were cultured in high-glucose Dulbecco’s modified Eagle medium (DMEM) (Meilun, Dalian, China), supplemented with 10% fetal bovine serum (FBS), under a humidified atmosphere at 37 ℃ with 5% CO₂.

Construction of a stable BCa cell line with ZWINT knockdown

Lentiviral vectors containing ZWINT knockdown and nonsense control sequences were purchased from Shanghai Genechem Company (Shanghai, China). The cells were treated with lentiviral infection solution as per the manufacturer’s instructions and cultured for 48 hours. Stably transfected cells were subsequently selected by culturing with puromycin for 2 weeks. Upon validation of knockdown efficiency, subsequent experimental procedures were initiated. The sequences of the short hairpin RNAs (shRNAs) used for ZWINT knockdown are listed in Table S2.

Quantitative reverse transcription polymerase chain reaction (qRT-PCR)

Total RNA was extracted from BCa cell lines with TRIzol reagent (Takara, Osaka, Japan) in accordance with the manufacturer’s protocol. RNA concentration and purity were assessed using a NanoDrop spectrophotometer (Thermo Fisher Scientific, Waltham, MA, USA). Complementary DNA (cDNA) synthesis was performed using the PrimeScript FAST RT Reagent Kit with gDNA Eraser (Takara). The SYBR Green assay (Takara) and the ABI 7500 real-time polymerase chain reaction (PCR) system (Applied Biosystems, Waltham, MA, USA) were employed to perform qRT-PCR. The following primer sequences were used: ZWINT, forward primer: 5'-CCTGGTTGAGTTTGTGGTGG-3', reverse primer: 5'-CTGTCGGCTCGTGTCTTCA-3'; glyceraldehyde-3-phosphate dehydrogenase (GAPDH) forward primer: 5'-ACAACTTTGGTATCGTGGAAGG-3', reverse primer: 5'-GCCATCACGCCACAGTTTC-3'.

Western blot

Cellular protein from BCa cells was extracted using radioimmunoprecipitation assay (RIPA) lysis buffer supplemented with 1% phenylmethylsulfonyl fluoride (PMSF) and 1% phosphatase inhibitor cocktail I (Meilun). Protein concentrations were determined using a bicinchoninic acid (BCA) assay kit (Solarbio, Beijing, China), and molecular weight markers (WJ103, Yeasen, Shanghai, China) were used to estimate the sizes of the target proteins. Equal amounts of protein were mixed with sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) loading buffer, boiled for 10 minutes, and separated via SDS-PAGE at 120 V for 1 hour. Proteins were transferred to PVDF membranes at 200 mA for 90 minutes. To allow individual detection of specific targets with different molecular weights, PVDF membranes were sectioned prior to primary antibody incubation. Target proteins included ZWINT (Abcam, Cambridge, UK), β-tubulin and p65 (Proteintech, Wuhan, China), and phosphorylated p65 (p-p65) and GAPDH (ZEN-BIOSCIENCE, Chengdu, China). PVDF membranes were blocked with a ready-to-use rapid blocking buffer (Solarbio) for 15 minutes at room temperature, followed by overnight incubation at 4 ℃ with primary antibodies diluted 1:1,000, followed by 1 hour incubation with horseradish peroxidase (HRP)-conjugated secondary antibodies diluted 1:10,000. Protein bands were visualized using an enhanced chemiluminescence (ECL) detection kit (Solarbio) and quantified with ImageJ software. As a result of this protocol, full-length membrane images with visible edges are not available. All uncropped and unprocessed original Western blot scans, including replicates, are provided as Appendix 1.

Cell Counting Kit-8 (CCK-8) cell proliferation assay

Cells were trypsinized and plated in 96-well plates, followed by incubation at 37 ℃ in a humidified incubator with 5% CO₂ for predetermined intervals (0, 24, 48, and 72 hours). CCK-8 solution (Meilun) was added to the wells, incubated for 1 hour, and the absorbance at 450 nm was read with a microplate reader.

Colony formation assay

Approximately 8×10² BCa cells were seeded into six-well plates and cultured for a period of 2 weeks. The cells were then fixed in 4% paraformaldehyde for 20 minutes and subsequently stained with 0.1% crystal violet for an additional 30 minutes. Clusters of more than 50 cells were considered as individual clones, and the number of colonies was quantified using ImageJ software.

Wound healing assay

In six-well plates where cells had formed a confluent monolayer, a sterile 200 µL pipette tip was employed to create a blank area. Following two washes with PBS, the cells were replenished with fresh serum-free culture medium. The cells were cultured in a 37 ℃ incubator with 5% CO₂ and observed under a microscope at 0 and 48 hours. The wound healing rate = [(0 hour − 24 hour)/0 hour] × 100%. The wound area was quantified using ImageJ software.

Migration and invasion assays

Cell migration and invasion assays were performed using Transwell chambers. In the lower chamber, culture medium containing 20% FBS was added. 5×10⁴ cells were seeded into the upper chamber. For migration assays, Matrigel was not used, whereas for invasion assays, the upper chamber was coated with Matrigel. The cells were incubated in serum-free culture medium at 37 ℃ in a 5% CO₂ atmosphere for 48 hours. After removing the small chamber, it was fixed in 4% paraformaldehyde for 20 minutes, followed by staining with 0.1% crystal violet for 30 minutes. Images of stained cells were captured under a light microscope, and the number of migrated/invaded cells was quantified using ImageJ software, with cell counts obtained from multiple randomly selected fields to ensure accuracy.

Cell cycle and apoptosis experiment

The cell cycle staining kit and Annexin V-allophycocyanin (APC)/7-aminoactinomycin D (7-AAD) apoptosis kit (LiankeBio, Hangzhou, China) were employed to analyze the cell cycle and apoptosis, respectively, following the manufacturer’s protocols. Stained cells (T24 or UM-UC-3) were examined using a flow cytometer (BD Biosciences, San Jose, CA, USA), and the data were analyzed using FlowJo software (version 10.8.1).

Transcriptome sequencing

Total RNA was extracted and submitted to Sangon Biotech (Shanghai, China) for RNA-seq. Library preparation, sequencing, and bioinformatics analysis were performed following the standardized protocols provided by Sangon Biotech. The processed gene expression matrix is provided in table available at https://cdn.amegroups.cn/static/public/10.21037tau-2025-710-1.xlsx.

Statistical analysis

In this study, statistical analyses were carried out using R (version 4.2.3), and all in vitro experiments were performed in at least three independent biological replicates, with each replicate measured in technical triplicate. When comparing continuous variables between the two groups, the Student’s t-test was applied for normally distributed data, whereas the Wilcoxon rank-sum test was used for non-normally distributed data. A P value of less than 0.05 (two-tailed) was deemed statistically significant.

Ethical statement

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Results

Differential gene analysis and the development and validation of a machine learning-based prognostic risk model

We utilized the edgeR package to identify DEGs in TCGA-BLCA (Figure 1A). Subsequently, 87 differentially expressed DNA repair-related genes in TCGA-BLCA were identified using Venn mapping (Figure 1B). By integrating 101 machine learning algorithms, we developed a DRRS based on nine prognostic genes identified through univariate Cox regression analysis (Figure 1C). The average C-index for each model was calculated across all datasets, and the Lasso + plsRcox model, with the highest overall average C-index of 0.649, was selected (Figure 1D). Using the regression coefficients from the plsRcox model, a risk score was calculated for each patient based on the expression levels of seven genes. Patients were then classified into high-risk or low-risk groups according to the median risk score (Figure 1E). Kaplan-Meier survival analysis demonstrated that patients with high DRRS had poorer progression-free interval (PFI) in TCGA (Figure 1F). In addition, the number of patient deaths increased steadily with higher risk scores (Figure 1G). Consistently, patients with high DRRS also showed worse overall survival (OS) in GSE32894 (Figure 1H). Univariate and multivariate analyses consistently validated DRRS as an independent prognostic indicator in BCa patients (Figure 1I,1J). Based on the ROC curve analysis, the areas under the curve (AUCs) of the DRRS were 0.635, 0.648, and 0.666 for the 1-, 2-, and 3-year follow-ups, demonstrating its reasonable prognostic performance (Figure 1K). Furthermore, the ROC and C-index curves revealed that DRRS exhibited superior predictive performance compared to other clinical traits (Figure 1L,1M).

Figure 1.

Construction of the DRRS using integrated machine learning techniques. (A) Volcano plots showing differential gene expression between normal and BCa tissues. (B) A total of 87 DNA repair-related genes showed altered expression between normal and BCa tissues. (C) The forest plot highlights 9 out of 87 intersecting genes with notable prognostic significance. (D) Machine learning was used to build multiple models, and the c-index was subsequently calculated for each model across the datasets. (E) Model genes and coefficients were determined by based on the optimal model. (F) Kaplan-Meier survival curve for PFI of patients in the low and high DRRS groups from the TCGA cohort. (G) Distribution of DRRS scores between low- and high-risk groups and OS status in the TCGA cohort. (H) Kaplan-Meier survival curve for OS of patients in the low and high DRRS groups from the GSE32894 cohort. (I,J) Univariate and multivariate Cox regression analyses of clinicopathological variables and DRRS scores in the TCGA cohort. (K) AUC values of DRRS at 1, 2, and 3 years in the TCGA cohort. (L) AUC values for DRRS and clinical characteristics. (M) The predictive performance of various clinical features, and DRRS was assessed using C-index curves. AUC, area under the curve; BCa, bladder cancer; BLCA, bladder urothelial carcinoma; C-index, concordance index; CI, confidence interval; DEGs, differentially expressed genes; DRRS, DNA repair-related risk score; Enet, elastic net; FC, fold change; GBM, generalized boosted regression modeling; Lasso, least absolute shrinkage and selection operator; OS, overall survival; PFI, progression-free interval; plsRcox, partial least squares regression for Cox; RSF, random survival forest; SuperPC, supervised principal components; survivalSVM, survival support vector machine; TCGA, The Cancer Genome Atlas.

Association of the DRRS with immune microenvironment, immune checkpoints, and therapeutic responses

To assess the involvement of DRRS in the tumor immune microenvironment of BCa, we evaluated its role in shaping immune cell infiltration across different DRRS groups. Based on seven immune infiltration algorithms, our results highlight a significantly higher degree of immune cell infiltration in the high-DRRS group (Figure 2A). Besides, we utilized the ssGSEA algorithm to analyze the differences in immune cell infiltration and immune function between the two groups. Compared to the low-DRRS group, the high-DRRS group exhibited elevated levels of CD8⁺ T cells, plasmacytoid dendritic cells (pDCs), regulatory T cells (Tregs), cytolytic activity, checkpoint activity, and APC co-inhibition (Figure 2B,2C). Significantly higher StromalScore and ESTIMATEScore were observed in the high-risk group, suggesting increased stromal components in the tumor microenvironment (TME). While ImmuneScore also showed a higher trend in the high-risk group, the difference lacked statistical significance (Figure 2D). Immune checkpoint-related gene analysis revealed significantly higher expression of PDCD1, CD274, LAG3, and CTLA4 in the high-DRRS group as opposed to the low-DRRS group (Figure 2E). To predict immune evasion in patients, we utilized the TIDE online tool, which demonstrated that higher TIDE scores were linked to a greater likelihood of immune evasion. Interestingly, although the TIDE algorithm predicted a higher tendency towards immune escape in the high-DRRS group (Figure 2F), real-world data from the IMvigor210 cohort demonstrated improved survival rates in this group following immunotherapy (Figure 2G). Additionally, the DRRS scores were higher in responders, and the high-DRRS group displayed a better response rate to immunotherapy (Figure 2H). Further analysis of chemotherapy response indicated that the high-DRRS group had inferior survival outcomes in contrast to the low-DRRS group (Figure 2I).

Figure 2.

Immunoinfiltration assessment. (A) Estimation of immune cell infiltration in the two DDRS subgroups using multiple algorithms. (B,C) Comparison of immune cell infiltration scores and immune-related function scores between the two DRRS subgroups. (D) Differences in TME scores between patients in different DRRS. (E) Comparison of immune checkpoint expression across different DRRS subgroups. (F) Prediction of immunotherapy response based on TIDE analysis. (G) Kaplan-Meier survival curves for OS in low and high DRRS groups from the IMvigor210 immunotherapy cohort. (H) DRRS scores of patients with CR/PR and SD/PD in the IMvigor210 immunotherapy cohort (top), and the proportion of immunotherapy responses in high and low DRRS groups within the same cohort (bottom). (I) Kaplan-Meier survival curves for OS in low and high DRRS groups from the GSE169455 chemotherapy cohort. The symbols *, **, and *** represent statistical significance as follows: P<0.05, P<0.01, and P<0.001, respectively. aDCs, activated dendritic cells; APC, antigen-presenting cell; CCR, CC chemokine receptor; CI, confidence interval; CR, complete response; DCs, dendritic cells; DRRS, DNA repair-related risk score; HLA, human leukocyte antigen; HR, hazard ratio; iDCs, immature dendritic cells; IFN, interferon; MHC, major histocompatibility complex; NK, natural killer; OS, overall survival; PD, progressive disease; pDCs, plasmacytoid dendritic cells; PR, partial response; SD, stable disease; Tfh, T follicular helper; Th, T helper; TIDE, Tumor Immune Dysfunction and Exclusion; TIL, tumor-infiltrating lymphocyte; TME, tumor microenvironment; Treg, regulatory T cell.

Differential expression and prognostic impact of ZWINT validated by GEO datasets and immunohistochemistry

To validate the expression of the representative gene ZWINT, we performed an analysis of external GEO datasets, comparing its expression across distinct sample types. Our analysis demonstrated that ZWINT was significantly upregulated in tumor samples relative to normal tissues (Figure 3A-3C and Figure S2A), with higher expression levels observed in recurrent cases compared to newly diagnosed patients (Figure 3D). Furthermore, in TCGA datasets, high ZWINT expression was significantly associated with unfavorable outcomes, including PFI and disease-specific survival (DSS) (Figure 3E and Figure S2B). Consistent with these findings, high ZWINT expression also correlated with poor DSS and OS in other datasets (Figure 3F-3J). To validate ZWINT protein expression, IHC staining was undertaken. As shown in Figure 3K, representative immunohistochemistry images revealed increased ZWINT staining in tumor tissues, which was statistically confirmed in Figure 3L, indicating significantly elevated protein levels compared to adjacent normal tissues.

Figure 3.

ZWINT expression and its prognostic significance in BCa. (A-D) Differential expression of ZWINT between normal and tumor tissues across TCGA-BLCA cohort, and multiple GEO datasets: GSE13507, GSE188715, and between primary and recurrent tumors in GSE120736. (E-J) Kaplan-Meier survival curves showing the prognostic value of ZWINT in various BCa cohorts. (K,L) IHC staining analysis of ZWINT protein expression in adjacent normal and tumor tissues (original magnification, ×200). The symbols * and *** represent statistical significance as follows: P<0.05 and P<0.001, respectively. BCa, bladder cancer; BLCA, bladder urothelial carcinoma; CI, confidence interval; DSS, disease-specific survival; GEO, Gene Expression Omnibus; HR, hazard ratio; IHC, immunohistochemical; OS, overall survival; PFI, progression-free interval; TCGA, The Cancer Genome Atlas; ZWINT, ZW10 interacting kinetochore protein.

ZWINT knockdown impairs proliferation, migration, invasion, cell cycle progression, and induces apoptosis in BCa

To further investigate the role of ZWINT in BCa, ZWINT was stably knocked down in T24 and UM-UC-3 cells using lentiviral transduction. Among the three designed shRNAs, shZWINT-2 was selected for subsequent experiments due to its consistent and robust knockdown efficiency at both messenger RNA (mRNA) and protein levels (Figure 4A,4B). The colony formation and CCK-8 cell proliferation assays demonstrated that ZWINT reduction significantly suppressed the proliferation of T24 and UM-UC-3 cells (Figure 4C,4D). Furthermore, ZWINT knockdown significantly decreased both the migratory and invasive capabilities of BCa cells, as shown by wound healing and Transwell assays (Figure 4E,4F). In addition, flow cytometry was used to explore the involvement of ZWINT in modulating the cell cycle and apoptosis of BCa cells. Analysis of cell cycle distribution revealed that ZWINT depletion triggered G1 phase arrest in both T24 and UM-UC-3 cells (Figure 5A). Similarly, ZWINT inhibition led to a marked accumulation of apoptotic cells in T24 and UM-UC-3 cells (Figure 5B). In summary, these findings demonstrate that decreasing ZWINT expression restricts the proliferative and metastatic potential of BCa cells, while facilitating cell cycle arrest and apoptosis.

Figure 4.

ZWINT knockdown suppresses proliferation, migration, and invasion in BCa. (A) qRT-PCR analysis of ZWINT mRNA levels in shZWINT and shctrl cells, normalized to GAPDH. (B) Western blot analysis showing ZWINT protein expression in shZWINT and shctrl cells, with β-tubulin as the loading control. (C) Colony formation assay was performed to assess cell proliferation, with representative images shown above and quantification below (cells were stained with crystal violet; images were acquired using a mobile phone camera). (D) CCK-8 assay to evaluate cell proliferation. (E) Wound healing assay reveals differences in migration (left), with statistical validation shown on the right (original magnification, ×100). (F) Transwell migration and invasion assay reveals differences in cell behavior (left), with statistical analysis on the right (cells were stained with crystal violet; original magnification, ×200). The symbols ns, *, **, and *** represent statistical significance as follows: not significant, P<0.05, P<0.01, and P<0.001, respectively. BCa, bladder cancer; CCK-8, Cell Counting Kit-8; GAPDH, glyceraldehyde-3-phosphate dehydrogenase; mRNA, messenger RNA; OD, optical density; qRT-PCR, quantitative reverse transcription polymerase chain reaction; sh, short hairpin RNA; ZWINT, ZW10 interacting kinetochore protein.

Figure 5.

Investigating the role of ZWINT silencing in modulating apoptosis and cell cycle in BCa cells. (A) The cell cycle assay indicates variations in cell cycle regulation (left), with statistical support shown on the right. (B) Apoptosis analysis by flow cytometry demonstrated differences in cell death rates (left), accompanied by statistical validation shown on the right. The symbol *** represents statistical significance as P<0.001. 7-AAD, 7-aminoactinomycin D; APC, allophycocyanin; BCa, bladder cancer; sh, short hairpin RNA; ZWINT, ZW10 interacting kinetochore protein.

Transcriptome sequencing and protein validation by Western blot

To gain deeper insights into the molecular mechanisms by which ZWINT modulates BCa cell behavior, we conducted transcriptome sequencing analysis on T24 cells subjected to ZWINT downregulation. Differential analysis was performed using the R package LIMMA, identifying 63 upregulated genes and 53 downregulated genes (|log₂FC| >1, P<0.05) (Figure 6A,6B). The functional enrichment analysis was performed to investigate the biological roles of the DEGs. GO enrichment analysis revealed distinct enrichment across the three major categories. For biological processes (BPs), the DEGs were primarily associated with epithelial cell proliferation, gland development, and the regulation of the ERK1 and ERK2 cascades. In the cellular component (CC) category, significant enrichment was observed in terms related to the endoplasmic reticulum lumen, intrinsic components of the synaptic membrane, and microtubule-associated complexes. Regarding molecular function (MF), the genes were involved in growth factor binding and peptide receptor activity (Figure 6C). KEGG enrichment analysis revealed that DEGs are primarily involved in key signaling pathways, such as the PI3K-Akt signaling pathway, Rap1 signaling pathway, and JAK-STAT signaling pathway (Figure 6D). In addition, GSEA was conducted to uncover pathways significantly enriched among the DEGs. Among the upregulated pathways, GSEA identified notable enrichment in the NF-κB signaling pathway, TNF signaling pathway, and the chemokine signaling pathway (Figure 6E). In contrast, the downregulated pathways included ribosome, phenylalanine metabolism, and ether lipid metabolism (Figure 6F). To further validate the involvement of these pathways, we conducted Western blot analysis on p65 and p-p65 in both control and ZWINT knockdown T24 and UM-UC-3 cells. As depicted in Figure 6G-6I, the ZWINT knockdown group displayed a significant increase in both p-p65 levels and the phosphorylation ratio of p-p65 to total p65 relative to the control group, while total p65 levels remained unchanged. These results suggest that ZWINT knockdown may be closely associated with the activation of the NF-kB pathway.

Figure 6.

Gene expression analysis comparing ZWINT knockdown versus shctrl groups in T24 cells. (A) Volcano plot depicting gene expression differences between shctrl and shZWINT in T24 cells. (B) The heatmap presents the top 50 genes with differential expression between shctrl and shZWINT in T24 cells. (C) GO enrichment analysis of DEGs in shctrl and shZWINT groups in T24 cells. (D) KEGG pathway analysis of DEGs between shctrl and shZWINT groups in T24 cells. (E,F) Through GSEA enrichment analysis, the most significantly upregulated and downregulated pathways were identified. (G) Western blot showing p-p65 and p65 expression in T24 and UM-UC-3 cells following ZWINT knockdown. (H) Quantification of p-p65 expression normalized to GAPDH. (I) Quantification of the p-p65/p65 ratio in T24 and UM-UC-3 cells. The symbols * and ** represent statistical significance as follows: P<0.05 and P<0.01, respectively. BP, biological process; CC, cellular component; DEG, differentially expressed gene; FC, fold change; GAPDH, glyceraldehyde-3-phosphate dehydrogenase; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; p-p65, phosphorylated p65; sh, short hairpin RNA; ZWINT, ZW10 interacting kinetochore protein.

Discussion

The molecular and clinical heterogeneity of BCa hinders the development of reliable biomarkers for early detection and personalized treatment, while also driving therapy resistance and complicating treatment efforts despite advancements in surgery, chemotherapy, and immunotherapy (17,18). Although cancer genomics has made substantial progress, the need for effective biomarkers remains unmet. Dysregulated DNA repair pathways are central to BCa progression and influence treatment outcomes, making them promising candidates for clinical biomarker development (19). To address these challenges, we constructed a seven-gene DRRS by integrating DEGs from TCGA-BLCA with DNA repair-associated genes, employing univariate regression analysis and 101 machine learning algorithms (16). DRRS demonstrated significant associations with survival outcomes, immune responses, and chemotherapy sensitivity. Additionally, ZWINT, selected as a representative gene within DRRS, was validated for its differential expression. Functional experiments revealed its role in regulating cell phenotype, apoptosis, and cell cycle progression, while transcriptomic analysis and Western blotting further suggested an association between ZWINT expression and NF-κB signaling activity.

Traditional staging systems fall short in predicting BCa prognosis and guiding therapy (20). To address this, we combined 10 machine learning algorithms validated for biomedical modeling into 101 algorithmic combinations for prognostic analysis. Through calculating the average C-index across the training and validation cohorts, we identified the combination of Lasso and plsRcox as having the highest average C-index. This approach resulted in the development of a seven-gene risk score, which we termed the DRRS. The results of Kaplan-Meier survival analysis, ROC curve, and univariate and multivariate Cox regression demonstrated the strong predictive capacity of DRRS across the training and validation cohorts.

DNA repair pathways, active in all components of the TME, influence cellular interactions and contribute to the remodeling of the immune microenvironment when dysregulated (11,18). Thus, we assessed the immune characteristics of patients stratified into high- and low-DRRS groups. In the high-DRRS group, key immune checkpoint-related genes, such as PDCD1 and CTLA4, were significantly upregulated. Intriguingly, TIDE analysis predicted a higher immune escape potential in this group, consistent with the observed upregulation of immune checkpoint genes. However, real-world data from an immunotherapy-treated urothelial carcinoma cohort revealed not only improved survival but also higher response rates in these patients, presenting an apparent contradiction. This may be explained by the elevated expression of immune checkpoint molecules, which potentially enhances responsiveness to ICIs by reactivating functionally exhausted CD8⁺ T cells (21-23). Additionally, ssGSEA analysis revealed that the high-DRRS group exhibited significantly enhanced activity of CD8⁺ cytotoxic T lymphocytes, dendritic cells, and T helper type 1 (Th1) cells, along with an upregulation of type I interferon responses. However, the downregulation of type II interferon responses and the concurrent increase in Tregs and immunosuppressive signals suggest a complex interplay of immune activation and suppression in the TME. Despite the higher level of immune cell infiltration, the significant upregulation of immune checkpoint genes indicates that the immune state remains suppressed.

Taken together, these findings suggest that the high-DRRS group exhibits characteristics of ‘cold tumors’, defined by an immunosuppressive microenvironment rather than a lack of immune cells (24,25). This highlights the dual nature of the high-DRRS group, where immunosuppressive signals coexist with a potentially responsive immune cell population. The DRRS model not only provides prognostic value but also holds promise for informing immunotherapy-related stratification. Further validation in larger cohorts and mechanistic studies is warranted to explore the interplay between immune escape mechanisms and therapeutic responses in high-DRRS tumors. Moreover, the prognostic potential of DRRS was evaluated in chemotherapy-treated patients. Results showed significantly worse survival in the high-DRRS group compared to the low-DRRS group, possibly linked to the DNA repair-related features of high DRRS, which may lessen the impact of chemotherapy-induced DNA damage (26,27). When patients are stratified by DRRS, a tentative treatment framework can be considered. High-DRRS tumors exhibit dysregulated DNA repair features together with immune infiltration and immune suppression, indicating a possible sensitivity to immunotherapy. Conversely, low-DRRS patients may experience greater benefit from chemotherapy due to relatively preserved DNA damage responses. Importantly, these implications are speculative and warrant confirmation in prospective trials.

ZWINT, selected as a representative gene in the DRRS model, demonstrated significant prognostic relevance across multiple datasets, including TCGA and GEO. Its expression was markedly elevated in tumor tissues compared to normal controls and further increased in recurrent tumors relative to newly diagnosed cases, indicating a potential role in tumor progression. IHC analysis confirmed these findings by demonstrating increased ZWINT protein levels in tumor samples, underscoring its importance as a prognostic biomarker and potential therapeutic target. Accordingly, ZWINT was selected for functional analysis based on its relevance within the model and its biological connection to cell cycle regulation and genomic stability.

Given the observed association between ZWINT expression and poor survival outcomes, we investigated its functional role in BCa progression. Knockdown of ZWINT in BCa cells led to significant reductions in proliferation, migration, and invasion, while inducing G1 phase cell cycle arrest and enhancing apoptosis. Mechanistically, ZWINT has been reported to regulate G1/S transition by modulating critical checkpoint proteins, including the downregulation of CDK4 and CCND1 (cyclin D1) and upregulation of the cell cycle inhibitor CDKN1B (p27) (28,29). Although these specific proteins were not directly assessed in our study, the observed G1 phase arrest aligns with previously reported mechanisms, suggesting a conserved role of ZWINT in cell cycle regulation across cancer types. In addition, ZWINT silencing significantly promoted apoptosis, highlighting its multifaceted role in maintaining cellular homeostasis. As a mitotic checkpoint gene, ZWINT is critical for chromosomal stability and proper cell division (30,31). Dysregulation of ZWINT has been implicated in tumorigenesis by driving uncontrolled proliferation and evasion of apoptosis. Consistent with these findings, studies in other cancers, including liver, lung, and pancreatic cancers, have demonstrated that ZWINT overexpression is associated with tumor progression and poor prognosis (29,32,33).

Transcriptomic profiling of ZWINT knockdown in T24 cells, coupled with functional enrichment analyses, revealed marked alterations in tumor-associated pathways and processes. GO analysis identified critical processes such as epithelial cell proliferation and chromosomal stability, essential for maintaining cellular integrity. KEGG pathway analysis highlighted dysregulation in canonical cancer pathways, including PI3K-Akt and JAK-STAT signaling. GSEA analysis revealed significant downregulation of pathways involved in cellular metabolism, such as ribosome biogenesis and phenylalanine metabolism. These changes suggest that ZWINT knockdown may disrupt key processes essential for tumor cell survival and proliferation, contributing to the observed inhibition of tumor growth. Conversely, inflammation-related pathways, including NF-κB signaling, were significantly upregulated. NF-κB signaling was among the pathways identified and was therefore further discussed given its known roles in immune regulation, inflammation, and tumor progression (34,35). Western blot analysis further confirmed increased phosphorylation of p65, indicative of NF-κB activation following ZWINT knockdown. This result initially seemed counterintuitive, as NF-κB activation is commonly associated with pro-tumorigenic effects, whereas ZWINT knockdown suppressed tumor cell proliferation and invasion in our study (36,37). Previous studies have reported that disruption of genomic stability can activate the p53/p21 pathway and lead to cell cycle arrest and apoptosis (32,33). In our study, ZWINT knockdown induced G1 arrest and apoptosis, suggesting a possible link to DNA damage-associated stress responses. Interestingly, this DNA damage response may indirectly trigger NF-κB activation, potentially as a compensatory mechanism to mitigate cellular stress (38,39). Moreover, the observed G1 phase arrest following ZWINT knockdown underscores the role of DNA damage in regulating cell cycle checkpoints. Prolonged or unresolved DNA damage during G1 phase could exacerbate cellular stress, further linking DNA damage repair pathways to inflammation-related processes such as NF-κB signaling (40,41). These findings highlight the complex interplay between DNA repair, cell cycle regulation, and inflammatory signaling, warranting further investigation into their roles in BCa progression.

Previous studies on BCa risk stratification have largely focused on prognostic performance alone. In this work, we integrated DNA repair-related gene features into a machine learning framework and validated the resulting risk score across multiple independent cohorts. In addition to prognosis, we explored the relationship between the DRRS and treatment response, including immunotherapy and chemotherapy. Functional investigation of ZWINT further extends the model by providing biological insight into a representative gene within the signature.

These findings emphasize the intricate relationship between DNA repair, cell cycle regulation, and inflammatory signaling in BCa progression, providing a foundation for further exploration. Future studies should focus on uncovering the molecular mechanisms through which ZWINT influences the TME and immune responses, potentially revealing new therapeutic avenues. Additionally, the clinical utility of DRRS and ZWINT warrants validation in larger and independent patient cohorts to ensure its applicability in clinical practice. Nonetheless, there are some limitations in this study. The functional experiments were limited to two BCa cell lines, which may not fully capture the heterogeneity of the disease. While transcriptomic analysis highlighted key pathways, we did not directly evaluate the expression of cell cycle-related proteins, leaving the precise molecular mechanisms underlying G1 phase arrest unconfirmed. Moreover, direct functional evaluation of DNA repair capacity was not performed. This limitation restricts a more definitive mechanistic interpretation of the role of ZWINT in DNA repair-related processes. In contrast, components of the NF-κB pathway were partially assessed, with increased phosphorylation of p65 serving as evidence of pathway activation. However, further studies are needed to comprehensively evaluate additional pathway components and their interplay with ZWINT. Additionally, the potential connection between ZWINT and immune modulation requires validation using in vivo models or patient-derived samples to clarify its role within the tumor immune microenvironment.

Conclusions

In conclusion, we developed DRRS using machine learning, which demonstrated strong prognostic value and potential for treatment stratification in BCa, including associations with immune response and chemotherapy sensitivity. ZWINT, selected as a representative gene within the DRRS model, was significantly associated with tumor progression and poor prognosis. Functional experiments revealed that ZWINT silencing suppressed tumor cell proliferation, migration, and invasion, induced G1 phase cell cycle arrest, and promoted apoptosis. Transcriptomic analysis identified NF-κB signaling as a key pathway altered by ZWINT knockdown, providing new insights into its involvement in inflammation and tumor-associated processes. These findings underscore the utility of DRRS and ZWINT as promising prognostic biomarkers and therapeutic targets, warranting further investigation in larger cohorts and mechanistic studies to fully elucidate their roles in BCa progression.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data Sharing Statement

Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-710/dss

tau-15-02-46-dss.pdf ^{(47.4KB, pdf)}

DOI: 10.21037/tau-2025-710